DSS 2026 — How to Model Cryptoeconomic Attacks and Adversaries in DeFi Abhimanyu Nag Source video: https://www.youtube.com/watch?v=kSmA9hMw4Wo Automatic captions — not human-reviewed. Complete spoken-audio coverage is unverified. Source language tag: en (English). Source export SHA256: 26b0df5c8af51d2468ca988d6d4fe20c26741f58d5a6c572e313c23fc39c1f40 [0:02] Hey everyone. [0:03] Full house. Nice to see that. [0:06] Uh [0:07] So, I'm Abhimanyu. Thank you so much for [0:10] coming to my talk. I'm talking about [0:12] cryptoeconomic attacks and adversaries [0:14] in DeFi. [0:16] This is going to be a more foundational [0:18] talk than some of the others. [0:21] This isn't, you know, solely, strictly [0:24] based on the MEV [0:26] literature that you all have seen or are [0:28] used to. And this is coming more from [0:30] the econ CS side from anything else. [0:32] I'm also my tax fellow at the University [0:34] of Alberta. [0:35] Would love to talk about it later. [0:39] So, [0:40] the main primitive that we're going to [0:42] be talking about here is going to be [0:46] auctions and how DeFi auctions have been [0:49] set up and how you can manipulate them. [0:53] There are a lot of examples of auctions [0:55] that you've seen in DeFi till now. [0:57] There's liquidation auctions, [1:00] trading auctions, which I mean as [1:01] intent-based auctions like cow swap, so [1:03] on and so forth. And there's also block [1:05] space auctions, which is [1:07] the normal [1:09] block space [1:11] allocation and payments that are [1:13] required by block producers. [1:15] So, fundamentally, an auction specifies [1:18] who can bid, [1:20] who receives what, and who pays. And the [1:23] main adversarial attack that I'm going [1:25] to be looking at for auctions today is [1:26] going to be sibyls where how can one [1:30] economic actor or a bidder or an agent [1:32] in a system is able to manipulate an [1:34] auction by spawning [1:37] multiple other identities. [1:39] So, [1:41] fundamentally, an auction has an [1:43] allocation rule and a payment rule. [1:45] And as as a very nice finite example, we [1:49] have two bidders, Alice and Bob. [1:52] So, Alice values a good at VA. [1:55] And so she bids BA [1:58] to be able to get a chance [2:00] gain that good. Similarly, you have Bob [2:03] who values that same good at VB bids VB [2:06] I guess. And the auction [2:09] specifies that exact function [2:12] which would basically be a probability [2:14] of [2:16] whether or not Alice or Bob will win the [2:18] auction for that good and how much [2:20] they'll have to pay for it. [2:22] There's a very simple example of a [2:23] single item auction which is going to be [2:25] specifically what I deal with today [2:28] which is going to be the second price [2:29] auction. So there's bids which are 10 7 [2:33] and 4. I bid for a single good at 10 7 [2:36] and 4. You see that the highest bidder [2:39] who bids 10 they end up paying the [2:40] second highest bid which is seven. [2:44] There are different DeFi auctions as I [2:46] talked about them. I'll just reiterate [2:48] the months. You've seen them for Maker [2:50] DAO which are collateral auctions which [2:52] is the significant example of a Dutch [2:55] auction where the collateral price [2:57] declines with time and purchases they [2:59] settle instantly. You have intent based [3:01] auctions on cow protocol where solvers [3:03] compete to propose settlements and then [3:06] on the more fundamental blocks place [3:08] level you see flash bots right and their [3:10] MV boost. [3:11] There are some adversarial questions [3:13] that will probably detract us from what [3:15] we I want to talk about today and I'm I [3:17] have limited time apparently so I'll go [3:19] over something more specific. [3:22] So an attack can exploit either code or [3:24] economics. That's the way I see it. Code [3:27] attacks are [3:28] all of you are probably way better at [3:30] that than I am. [3:31] And so today I'm going to be [3:32] specifically looking at economic [3:34] attacks. [3:35] Some of the examples of economic attacks [3:37] right now are as I said sibyls [3:40] which is one person spawning multiple [3:42] agents. At the same time there's [3:43] collusion attacks where it's all of the [3:46] agents in the system collude with each [3:48] other. There's a really subtle [3:48] difference with sibyls. [3:50] It's basically that these these are [3:51] separate identities right. [3:53] So the very basic model of an attacker [3:56] is that exact action which maximizes [3:59] their utility [4:01] given the information constraints that [4:02] have, and you also minimize the cost [4:06] that you need to actually conduct the [4:08] attack. [4:11] Right now, some specific assumptions [4:14] [snorts] that we're going to be making [4:15] for auctions is going to be that it's [4:18] truthful, where if you remember from our [4:20] example, there's VA and BA. [4:24] Truthfulness implies that VA is going to [4:25] be BA no matter what because that's the [4:27] highest utility [4:29] for the agent. [4:31] At the same time, civil proofness where [4:32] as I said, one actor, they cannot gain [4:34] more utility by dividing into several [4:37] bits. [4:38] And the main point I want to drive is [4:39] that if an auction is considered safe [4:41] for one specific bid, it might not be [4:43] this it might not be safe for multiple [4:45] bids at the same time. [4:47] So complete civil attack, very very [4:49] simple example. Think of a lottery, [4:51] right? [4:52] You have a prize which is say worth 120 [4:54] and there is a lottery here where your [4:57] probability of getting the item is [4:59] proportional to the number of players in [5:01] the game. [5:02] Say [5:03] if it's 120, then the probability of you [5:06] getting and if sorry, if there are say n [5:08] bidders, your probability of winning is [5:10] one by n, right? So let's have a [5:13] baseline of three identities. [5:15] I have an attacker and I have two honest [5:18] people, right, who are bidding alongside [5:19] me. [5:21] I my way of attacking would be to divide [5:24] my one bid into two or maybe, you know, [5:26] I don't have to divide that one value. I [5:28] can have a value greater than that so on [5:29] and so forth. The main thing that you [5:31] have to think about is that when there's [5:32] one attacker with two identities, it's [5:34] my probability of winning the good is [5:36] one by three, which is $40 in this case. [5:39] Similarly, if my once I deviate and I [5:42] divide my bid into two, my actual [5:45] probability of winning is one half now, [5:47] which is $60. So you see that difference [5:50] of $20 there, right? [5:52] This will give you some more intuition [5:54] where you might think that in order to [5:56] make the system or the auction [5:58] completely secure, you're going to have [5:59] to impose a certain cost which has to be [6:02] at least greater than $20 in this case. [6:05] So now that you think about it, when you [6:07] take it across agents, across people, [6:09] attack search is an optimization [6:10] problem. [6:11] And what we are actually going to be [6:13] optimizing is that the utility you get [6:15] from being a civil agent versus a [6:17] truthful agent, you're basically trying [6:19] to maximize that exact portion, which is [6:21] delta here as you see it on the screen. [6:25] There's several ways to calculate it, [6:26] which are grades, normal optimization [6:29] techniques that you see in Python. [6:31] There's also some constraints here with [6:33] gas, deposits, timings. You get the [6:35] gist. [6:36] The main point right now, as of today, [6:39] the second price auction is the only [6:41] auction that's proven to be civil proof [6:43] for now in the single item case. There's [6:45] a lot of literature in the multi item [6:47] case as well that's been done by the [6:49] wonderful people at Cow Protocol, right? [6:51] You should definitely check them out. [6:53] We have some more results for [6:56] civil proof auctions. Maybe I'll talk [6:58] about it later once we publish the [6:59] paper. [7:01] So one really interesting aspect that I [7:02] want to talk about is AI agents as [7:05] attackers. [7:07] So [7:08] AI agents are the way I see it on a pure [7:11] economic level, they're not going to be [7:12] very different than human beings about [7:13] thinking, [7:15] about how to manipulate a protocol. [7:18] They're going to do the same exact thing [7:19] that we do. They're going to read the [7:20] rules. They're going to form a [7:21] hypothesis about how to actually attack [7:23] it. They're going to evaluate how much [7:25] it's going to cost them to actually [7:27] conduct the attack and they're going to [7:29] basically revise their constraints from [7:32] feedback, market feedback, people [7:33] feedback, so on and so forth. [7:37] The next thing that I want to talk about [7:39] is a very, very simple [7:41] infrastructure [7:42] for how you as an agent or like you [7:45] having an agent, please don't do this. [7:47] Your AWS is is going to hate you. Your [7:49] bills are going to be huge. I'm not [7:51] going to lie. [7:52] But the main thing that I want to talk [7:53] about is how [7:56] what a basic infrastructure might look [7:58] like, right? [8:00] So I have been an AWS max here all my [8:02] life. I'm sorry if I offend someone here [8:04] with that, but I would basically take an [8:08] S3 container where I'm going to query [8:11] information or data from Dune or [8:14] on-chain data. Hopefully you have a node [8:15] setup of your own. [8:17] You're going to take the on-chain data, [8:18] you're going to take all the code, the [8:20] hash, the seeds, [8:22] and you're going to give it to your [8:23] agent. Your agent after pulling it is [8:25] going to form a hypothesis, right? It's [8:27] going to take the information at a [8:29] microscopic level. It's going to skim [8:31] through it, figure everything out. It's [8:33] going to give you a structured point of [8:35] attack. [8:37] Then in order to solve the deviation, in [8:40] order to actually see all the values, [8:42] all the bits that it has to do, how much [8:45] civil bit do they actually need to [8:47] actually manipulate the protocol, [8:48] they're going to use [8:50] an EC2 worker. That's what I would [8:52] presume. That's what I would do [8:53] technically, where I would do [8:54] computations at [8:56] at the millisecond level, I presume. [8:59] I I don't think AWS does that. It's [9:01] pretty slow. [9:02] But on top of that, once those [9:04] simulations are complete, you're going [9:05] to know the agent at least is going to [9:06] know by the end of it exactly what the [9:09] deviations are required. [9:10] Then there's going to be an independent [9:12] replay, which is again like that's big [9:15] words for saying back testing where [9:17] you're going to take block data, you're [9:19] going to see the civil deviations that [9:21] would have given you more rewards than a [9:22] normal person. Then you can back test [9:25] it. [9:26] You can see exactly what gain, what [9:29] utility you're going to get as an [9:31] attacker versus a normal agent. And then [9:33] you're going to verify it, [9:35] send it back to S3 as logs, right? You [9:38] can pull it, you can check it, verify [9:40] it, so on. [9:42] So I am not going to leave you with just [9:45] like you know just that, just a way to [9:47] attack the system. I'll also talk about [9:50] I'm I just going to want to talk about [9:51] how [9:52] an sorry, how a protocol designer is [9:55] able to figure out a mechanism or learn [9:58] a mechanism against an attacker. [10:00] For that, I the way I see it, it's going [10:03] to be a coupled optimization function. [10:07] Big words, but what it actually is is [10:09] that you're going to have an inner [10:10] problem where you're going to freeze the [10:12] mechanism parameters [10:14] and you're going to search for Sybil [10:15] attacks for any auction. You have an [10:17] auction A with payment rule A payment [10:19] rule B sorry and you're going to figure [10:21] out a way [10:23] you're going to put a million Sybil bids [10:26] to make as many bidders as possible from [10:28] one person and see what your utilities [10:31] are in that case [10:35] and then you're going to solve the outer [10:36] problem where you're going to solve it [10:38] for the protocol designer now. The outer [10:40] problem is basically I want to increase [10:43] I want to maximize my revenue at the [10:45] same time minimize [10:46] the instances of Sybil's. [10:49] What I mean by that is that when you [10:51] minimize the inner problem [10:53] you're basically figuring out exactly [10:56] how [10:57] expensive it is for a Sybil attacker to [10:59] actually attack the protocol. You're [11:01] going to increase that expense sorry [11:03] you're going to make that really really [11:04] expensive as expensive as you can while [11:06] at the same time [11:08] improving your revenue maximizing your [11:10] revenue at that very instance. What I [11:13] mean by that again at the end is that [11:17] you're going to end up learning a [11:18] mechanism where you know using neural [11:21] networks deep learning that's what we [11:22] did for our very recent paper [11:25] hopefully when when it comes out I'll be [11:27] able to talk more about it but what it [11:29] actually does is it's able to learn [11:31] mechanisms where attackers no matter [11:34] what they do do not have any incentive [11:37] to attack the auction at least using [11:39] Sybil's. [11:41] At the end of it I have some final [11:43] thoughts I have like three minutes left [11:45] as well [11:45] that [11:48] the full proforma [11:50] of actually conducting a Sybil attack or [11:53] thinking about attackers in [11:55] economic protocols in general is you [11:57] have to specify the agents you have to [11:59] specify exactly how Sybil bids can be [12:02] inputted into the system you have to [12:04] compare the utilities between being an [12:05] attacker and being a normal agent and [12:08] you have to separate when I say evidence [12:10] from guarantees what I actually mean is [12:12] that separate back-tested variables with [12:15] the actual market conditions. [12:17] And my final point that I want to drive [12:18] with this [12:19] small talk, I guess, [12:21] not so small now that I think about it, [12:23] is that economic security it actually [12:25] needs an adversarial model. It doesn't [12:27] need like a specific it doesn't need [12:29] like a best-case model to be fair. [12:32] The worst-case scenarios are always [12:33] considered in consensus research. It's [12:36] always [12:37] considered in even in auction [12:39] literature. But I I haven't I'm yet to [12:41] see ways in which people are able to [12:44] combine those two. And by finally, I [12:47] just want to say that I we have figured [12:49] out a way to create civil-proof auctions [12:51] from the ground up, turn any auction [12:54] into a civil-proof one using [12:56] wistfulness. So that'd be something I'd [12:59] love to talk about. [13:01] That's all I had. If you have questions, [13:03] please come find me. It was great to [13:04] meet you. Thank you so much.